Base build recipe
This describes the base release. Incremental additions are appended without resampling it; current counts and per-source transformations are recorded in the release audit and addition manifests.
- Canonical recasts. Each Tasksource task is converted deterministically.
- Criteria are the source's own label names and answer options.
- Multiple-choice rows keep every option in a per-row order.
- A final "all/none of the above" reads "all/none of the other options".
- Options that cite other options by letter or number keep their order.
- The question is the task's own when its inputs alone do not say what to predict ("What stance does the tweet take on feminism?"), and a generic instruction otherwise. Label-verification and packed questions carry it too.
- Variants. Low-frequency, deterministic variants cover label verification as
noul, criterion order, and instruction wording. - Packing. Up to 10% of each classification task's examples are packed, two to four at a time, into
packed_derivedstates. Their questions (an item's label, agreement, existence, counts) follow exactly from the gold labels. - Mixing (
full). Formats get fixed shares of the train rows (47% classification, 30% multiple choice, 3% token labeling, 10% graded (soft-label sources), 10% procedural). Within a format, dataset families get equal shares, scaled by hand-set weights (more for adversarial NLI, long documents and preference pairs; less for templated probes), times audit weights from a per-task check of Jev on 200 examples: ×1.5 for hard tasks whose gold is right by construction (synthetic logic, theory of mind, spatial reasoning), ×0.5 for near-solved tasks and for hard tasks whose gold is a judgment call (ratings, preferences, crowd sentiment). Sources with many options get slightly more room. Related questions are kept together. - Mixing (
default). About 1M rows drawn fromfull, never repeating a row. Buckets of related sources get set shares (jev_mixes.py: 15% logic, 10% NLI, 10% knowledge QA, 9% long documents and fact-checking, 8% intent and routing, 12% procedural, ...). Inside a bucket, sources get rows by score × √size. The score multiplies:- correctness: sources that failed label review get 0;
- the zone of proximal development: how much probability the decision models (Jev, Liquid D1) give the gold answer. Near-solved sources and sources the models miss outright both get less;
- interest and cleanliness, from Jev yes/no checks for transferable skills, trivial examples and malformed rows;
- ×0.6 for two-option tasks (about a third of the rows);
- ×2 for sources picked by reading them.
- Order and coverage.
- The first 1,000 train rows are interleaved to show variety in the Dataset Viewer; the rest is shuffled. Questions of a group stay adjacent throughout.
- Evaluation benchmarks (BIG-bench, MMLU, BLiMP, MATH test, ...) are left out so they stay clean for evaluation.
- Sources. sources.yaml lists every source with its rows, the Hub dataset and revision it was loaded from, the original dataset behind each tasksource copy, and its licenses.
- Audit trail. The source mix, failed source list, and build manifest ship with the data.
- Reproducible. The build runbook rebuilds the release from Tasksource's task catalog.